Spectral models are still the mainstream in global operation NWP now. Based on global spectral model, we discuss two-time-level semi-Lagrangian and semi-implicit integration scheme and quasi-cubic spatial interpolation for the terms evaluated at the departure point. A on-demand massage-passing scalable parallel algorithm has been developed. The algorithm is examined in a 4 nodes dual-CPU SMP Linux cluster. The relative speedup of 4 tasks to 8 tasks is 1.65.